A Study on the Association Between Tower Crane Operator Fatigue State and Collision Risk Under Human–Machine Interaction

Abstract

To investigate the relationship between operator fatigue and collision risk under human–machine interaction (HMI) in intelligent tower crane operations, and to reveal the mitigating effects of HMI on fatigue-induced collision risks, a comprehensive data acquisition approach integrating eye-tracking signals, risk indicators, and fatigue scale assessments was proposed and validated through scenario-based experiments. First, two experimental scenarios—traditional mechanical operation and HMI operation—were established. Based on a review of existing studies, representative eye-movement metrics and fatigue scale indicators were selected. Subsequently, operator fatigue states were classified into three levels: low fatigue, moderate fatigue, and high fatigue. A total of 28 participants were recruited to complete fatigue assessments and subsequently perform tower crane lifting tasks under both experimental scenarios. Finally, collision risk under different scenarios was quantitatively evaluated using the safety distance between the crane hook and the rigger, as well as the frequency of collision alarms. The results indicate that, under traditional mechanical operation, increasing fatigue levels were associated with a significant reduction in safety distance between the crane hook and the rigger, accompanied by a marked increase in collision alarm occurrences, resulting in a relatively high overall collision risk. In contrast, under the HMI operation scenario, participants demonstrated superior operational control at equivalent fatigue levels. Specifically, under moderate fatigue, collision risk was reduced from low risk to no risk, while under high fatigue, collision risk decreased from high risk to low risk. These results indicate that, under laboratory-simulated conditions, human–machine interaction can mitigate, to a certain extent, the increasing trend of collision risk when operators perform tower crane lifting operations under fatigue. These findings provide a scientific basis for further optimization of intelligent tower crane operational modes and the development of enhanced safety management strategies.

1. Introduction

Tower cranes represent essential large-scale lifting equipment on construction sites, and their operation depends primarily on manual control. Tower crane operation tasks are typically characterized by long working durations, high operational intensity, and complex environments, placing substantial demands on operators’ attention, reaction capability, and operational precision. Under high-intensity and high-workload conditions, operators are prone to developing both physical and cognitive fatigue, which can significantly degrade operational performance and consequently increase the risk of safety incidents such as collisions. Therefore, exploring effective approaches to mitigate the adverse effects of fatigue on tower crane operators’ operational performance in complex construction environments, and to enhance the overall safety level of lifting operations, has gradually emerged as a critical research topic in the field of construction safety management. In this context, HMI technologies, with advantages such as real-time feedback, intelligent assistance, and corrective control, provide a promising technical approach for improving operational accuracy and reducing human error, and are expected to play a key safety-regulating role under fatigue conditions.

Against this background, it is imperative to conduct a systematic investigation into the effects of operator fatigue on collision risk during tower crane operations, as well as the regulatory influence of HMI technologies on operational accuracy and collision risk under fatigue conditions. Elucidating the intrinsic relationships among operator fatigue, operational behaviors, and collision risk, as well as uncovering the mechanisms through which HMI exerts its regulatory effects, can not only enhance the safety of tower crane operations but also provide theoretical foundations and technical support for the advancement of intelligent and smart construction equipment.

Extensive research has been conducted on the relationship between operator fatigue and construction safety risk. Existing studies indicate that fatigue significantly impairs operators’ attention levels, reaction speed, and decision-making ability, thereby increasing the likelihood of collisions and accidents [1]. Most related studies have investigated fatigue mechanisms from behavioral and cognitive perspectives, demonstrating that fatigue leads to extended reaction times, reduced attentional resource allocation, and increased cognitive workload, which in turn increases operators’ susceptibility to operational errors during complex tasks [2,3]. In addition, physical fatigue not only directly promotes unsafe behaviors among construction workers but may also further amplify safety risks through mediating factors such as job burnout. Meanwhile, fatigue monitoring methods based on physiological indicators—including heart rate, eye-movement parameters, and electroencephalogram (EEG) signals—have been extensively applied in fatigue assessment and risk prediction studies, providing important technical support for the objective identification and quantitative analysis of fatigue states [4].

In the field of human–machine interaction, HMI technologies have found extensive applications in transportation, industrial automation, and intelligent manufacturing. Intelligent driving assistance systems improve traffic safety by providing environmental perception, risk warning, and control assistance functionalities. In industrial applications, various HMI modalities—including touch, voice, and gesture interactions—combined with the integration of virtual reality (VR) and augmented reality (AR) technologies—have significantly enhanced user experience and reduced operational errors [5,6]. Particularly in industrial and construction robotics, multimodal HMI analysis frameworks integrating operational data, physiological signals, and environmental information have offered novel insights into intelligent decision-making and collaborative control [7]. However, existing studies have predominantly focused on scenarios such as traffic driving, aviation control, and industrial automation, where investigations into the mechanisms underlying fatigue and HMI are generally conducted in relatively closed or standardized operational environments. In contrast, tower crane operations are characterized by high-altitude tasks, dynamic multi-target interference, significant three-dimensional collision risks, and multi-role collaborative coordination. The risk structure of such operations differs substantially from that of driving or conventional industrial control contexts. In high-risk construction scenarios of this nature, operators’ judgments of spatial distance rely heavily on visual perception and cognitive resources, and fatigue may exert a more sensitive influence on their spatial decision-making capability. Although prior research has separately examined the effects of fatigue on safety behavior and the supportive role of HMI technologies in operational safety, systematic experimental investigations that integrate fatigue states with intelligent HMI mechanisms in the context of tower crane operations remain limited. In particular, empirical evidence is lacking regarding whether HMI can moderate collision risk indicators under different levels of fatigue and how the boundaries of such moderating effects can be defined.

Against this background, the present study focuses on the operational context of tower cranes and establishes an integrated analytical framework encompassing fatigue, collision risk, and HMI intervention. An experimental simulation study was conducted accordingly. The main contributions of this research are threefold. First, an analytical model describing the relationship between fatigue state and collision risk in tower crane operations is developed, thereby extending the application of fatigue research to high-risk construction equipment contexts. Second, experimental simulations are employed to quantify variations in safety distance control and risk exposure characteristics under different fatigue levels, revealing the influence of fatigue on spatial risk judgment. Third, the moderating trend of HMI on collision risk indicators under varying fatigue conditions is examined, providing empirical support for the design and implementation of intelligent tower crane assistance systems. Therefore, this study not only applies the existing fatigue–HMI relationship to a specific operational context, but also conducts a contextualized validation and structural integration of their interaction mechanism within high-risk construction equipment operations.

Based on the proposed framework, an intelligent tower crane operation platform was established in a laboratory setting. Two experimental scenarios were designed, namely traditional manual mechanical operation and HMI—assisted operation, with three fatigue levels (low, moderate, and high) configured for each scenario. Fatigue levels were comprehensively assessed through a combination of physically induced tasks, subjective rating scales, and eye-tracking indicators. Operational process data were recorded using displacement sensors. The analysis was conducted at three levels—fatigue state, operational behavior characteristics, and collision risk indicators—to examine the moderating effect of HMI on risk trends under different fatigue conditions. This study was conducted under controlled laboratory conditions. Accordingly, the findings should be interpreted as an exploratory analysis of the interaction between fatigue and HMI, and their applicability to real construction environments requires further validation.

The remainder of this paper is organized as follows. Section 1 presents the introduction, highlighting the safety risks associated with tower crane operator fatigue and the regulatory potential of HMI. Section 2 reviews the relevant literature and establishes the research foundation. Section 3 describes the methodology, including experimental design, physical fatigue induction, and data analysis. Section 4 introduces the experimental participants, experimental scenarios, and experimental procedures. Section 5 presents the results, including descriptive statistics, repeated-measures analysis of variance, linear mixed-effects model analysis, and collision risk analysis under the two operational scenarios. Section 6 discusses the interpretation, implications, and practical significance of the findings. Finally, Section 7 concludes the study, emphasizing the mitigating effects of HMI on operator workload and collision risk, and discusses the limitations of the present research.

2. Literature Review

2.1. Effects of Operator Physical Fatigue on Construction Safety Risk

Physical fatigue, as one of the most prevalent yet concealed risk factors on construction sites, has been widely demonstrated to significantly increase construction safety risks through multiple pathways. Its generation mechanisms are typically associated with high-intensity, prolonged, or repetitive physical work, which leads to the accumulation of skeletal muscle metabolites, reductions in muscle strength and coordination, and subsequent destabilization of the neuromuscular and information-processing systems. An increasing body of research suggests that physical fatigue may not only impair workers’ physical performance, but also influence their perception, judgment, and behavioral responses to construction-related risks, thereby potentially affecting the process of accident occurrence [8,9].

Physical fatigue substantially impairs operators’ hazard recognition capability and overall situational awareness. Both experimental and field studies consistently show that fatigued workers display a significantly narrowed visual search range, with attention more readily directed toward non-critical areas, resulting in delayed detection or omission of potential hazard cues [10,11]. Eye-tracking analyses conducted in virtual construction simulation environments further demonstrate that, with accumulating fatigue, both the accuracy and efficiency of hazard recognition decline simultaneously. Questionnaire and experiment-based studies have further confirmed a significant negative correlation between physical fatigue and hazard identification as well as safety risk perception, revealing that physical workload undermines workers’ situational awareness by disrupting attention allocation and visual processing, thereby elevating accident risk [12]. Moreover, physical fatigue alters operators’ behavioral decision-making preferences and induces more unsafe behaviors. When operating under high fatigue levels, individuals tend to adopt “effort-saving” or “time-saving” strategies to compensate for reduced physical capacity; however, such strategies are frequently associated with elevated risk and reduced safety compliance [13]. Research also indicates that when physical fatigue is combined with cognitive fatigue, it manifests not only as attentional decline and delayed judgment but also as unstable gait, reduced grip precision, and increased movement variability, thereby elevating the likelihood of operational errors, slips, and falls [13,14,15]. These findings align with field investigation results, which identify physical fatigue as a prevalent yet often overlooked contributor to construction accidents in safety management practices [16].

Meanwhile, physiological monitoring provides objective support for the aforementioned mechanisms. Multisource physiological signal studies indicate that indicators such as heart rate and heart rate variability, skin temperature, muscle activity, and physical workload display consistent and identifiable patterns under fatigue conditions and are statistically associated with the occurrence of unsafe behaviors [17,18,19]. In construction and industrial work contexts, integrating wearable sensors with task-context models—considering work intensity, environmental load, and cumulative time effects—can substantially enhance fatigue detection, reduce missed detections under high-risk conditions, and provide a foundation for proactive safety warning systems [20]. Furthermore, time-series modeling approaches, such as deep learning methods, enable the extraction of dynamic features from physiological signals, allowing fatigue state identification to evolve from static assessment toward real-time prediction [21].

It should be noted that physical fatigue rarely acts in isolation but instead interacts with psychological fatigue, environmental conditions, and organizational factors to influence construction safety outcomes. Studies have shown that long working hours, high-temperature or low-illumination environments, repetitive tasks, and night-shift schedules can all amplify the cumulative effects of physical fatigue [22]. For example, under low-light conditions, the combined effects of physical fatigue and visual load further impair postural control and spatial judgment, significantly increasing collision risk [23]. In addition, physical fatigue often increases accident probability through indirect pathways by influencing risk attitudes and safety behavioral intentions, indicating that it should be considered a systemic, cross-level safety risk factor requiring coordinated management alongside engineering controls, work organization, and safety culture initiatives [24]. Further research has indicated that construction safety risks exhibit cascading and networked characteristics, whereby different risk factors may interact through associative pathways and generate amplification effects [25]. From this structural perspective, fatigue should not be regarded solely as an individual physiological condition; rather, it may be understood as a constituent element within the risk network, influencing overall safety performance through its interactions with environmental and behavioral factors. Therefore, fatigue management should be coordinated with engineering controls, work organization, and safety culture development to enhance system-level safety performance.

2.2. Research on Collision Risk in Intelligent Tower Crane Operations

As critical heavy equipment in building construction, tower cranes involve substantial safety risks during operation, particularly in multi-crane overlapping work zones and complex construction environments, where collision risk becomes a major factor affecting both construction safety and operational efficiency. In recent years, with the advancement of intelligent and digital technologies, researchers have conducted extensive investigations into collision risks in intelligent tower crane operations, covering multiple aspects such as multi-objective scheduling optimization, path planning, dynamic collision warning, and digital safety management. To address collision risks in multi-crane overlapping work areas, existing studies have proposed multi-objective optimization models that coordinate task allocation and operation sequences among multiple tower cranes to reduce collision probability and energy consumption costs. Some studies have unified task selection and lifting sequence into a single optimization framework and applied the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) to rank Pareto-optimal solutions, achieving coordinated collision avoidance and cost optimization [26]. Meanwhile, a Cooperative Co-evolutionary Genetic Algorithm (CCGA) was introduced to enable collaborative allocation and priority scheduling of tasks within overlapping operational areas. This approach aims to ensure that, as far as possible, tasks within the same area and time period are executed by a single crane, thereby reducing potential collision risks to a certain extent. The proposed method was further applied to a large-scale airport construction project, where its performance was analyzed and evaluated [27]. In addition, scheduling models based on mixed-integer linear programming (MILP) have enabled collision-free operations in multi-crane overlapping zones while simultaneously considering transportation efficiency and safety constraints [28].

In terms of dynamic path planning, research focus has gradually shifted toward real-time collision avoidance methods that integrate computer vision, point cloud data, and building information modeling (BIM) technologies. Studies have developed dynamic collision warning mechanisms based on visual perception and trajectory prediction, using multi-object tracking algorithms to obtain time-series positional information of workers and loads, and employing deep learning models to predict future trajectories for dynamic risk level assessment [29]. A dynamic path planning model developed based on video data and instance segmentation techniques was proposed to mitigate safety threats posed by lifting operations to ground personnel. Through path risk assessment and optimization, the model was shown to reduce potential hazards to on-site workers to a certain extent [30]. Furthermore, automatic lifting paths generated using environmental point cloud data, combined with octree sampling strategies and improved A* algorithms, have enabled efficient generation of executable paths, reducing path collisions and operational complexity [31]. In dynamic BIM environments, automatic re-planning methods for lifting paths have achieved near real-time collision avoidance and path optimization [32].

In the field of construction monitoring and collision detection, the integrated application of BIM and unmanned aerial vehicle (UAV) technologies has received increasing attention. By constructing three-dimensional construction scenes, crane path simulation and collision detection can be performed, and potential safety hazards can be eliminated through crane position adjustment [33]. For high-rise modular integrated construction, collision detection strategies combining octree-based spatial partitioning and bounding box algorithms have been proposed, significantly improving computational efficiency while maintaining detection accuracy [34]. In addition, minimum-distance calculation methods based on oriented bounding boxes provide quantitative support for real-time collision warning during lifting operations [35].

With respect to crane layout planning and optimization of overlapping work zones, related studies have employed mixed-integer linear programming models to jointly optimize the number, type, location of tower cranes, and supply points, thereby minimizing overlapping interference and enhancing operational safety [36]. Agent-based modeling approaches for dynamic supply point selection have achieved optimal matching between tasks and supply points, improving efficiency and safety performance in overlapping areas [37]. Meanwhile, the crane operation sequence was modeled using the Multiple Traveling Salesman Problem (MTSP), which enabled compliance with task priority constraints while reducing collision risks to a certain extent [38]. In addition, simulated annealing algorithms have been applied to spatiotemporal planning of tower cranes, enabling coordinated operations and collision avoidance in overlapping areas across different construction stages [39]. Furthermore, the application of artificial intelligence and deep learning technologies in collision risk identification and warning has continued to expand. Deep learning-based object detection and distance measurement systems, integrating visual and sensor data, have enabled dynamic collision prevention during lifting operations [40]. Transfer learning methods have been employed to identify unsafe lifting behaviors such as crane tilting and sudden unloading, providing data support for proactive risk management [41]. Additionally, systematic reviews of digital safety management technologies for tower cranes have provided theoretical foundations for collision risk control from both quantitative and qualitative perspectives [42].

2.3. Role of HMI in Construction Collision Risk

As the construction industry continues to evolve toward intelligence, digitalization, and automation, construction sites are increasingly characterized by highly integrated equipment, constrained working spaces, and frequent interactions between workers and machinery. Consequently, construction collision risks have become more complex, exhibiting dynamic, concealed, and systemic characteristics simultaneously. In this context, HMI is widely regarded as a critical link connecting construction personnel, intelligent systems, and the working environment, playing an indispensable role in the identification, assessment, and intervention of construction collision risks. Existing studies indicate that construction collision accidents are rarely caused by a single technical failure or human error, but rather stem from systemic failures resulting from insufficient information exchange or cognitive inconsistency among humans, machines, and the environment [43]. Therefore, optimizing HMI to enhance the perceptibility, interpretability, and responsiveness of risk information has become a key research direction in construction safety.

At the early stage of construction collision risk identification, HMI primarily relies on multisource sensing and information visualization mechanisms to enhance workers’ perception of potential hazards. Bibliometric studies have shown that the core value of HMI in construction hazard identification lies in integrating computer vision, sensing technologies, and intuitive interactive interfaces, enabling timely perception and recognition of latent hazard states [44]. The CHR-HCI framework proposed by some scholars further demonstrates that human-centered human–computer interaction design can, to a certain extent, effectively bridge the gap between automated recognition outcomes and construction workers’ risk perception [45]. In recent years, vision-based HMI systems have been widely applied in lifting equipment operations and other high-risk scenarios. By dynamically annotating worker positions, hazardous zones, and motion trajectories on monitoring interfaces, these systems have significantly reduced construction collision risks [46]. Meanwhile, interaction models for human–machine collaboration have gradually expanded from single-equipment monitoring to comprehensive perception of workers’ behaviors and postures. Studies based on human skeleton recognition and interactive feedback have enabled real-time understanding of worker activity states, providing situational awareness foundations for collision risk warning [47].

Traditional construction risk assessment results are often presented in abstract indicators or textual forms, which are difficult for frontline workers to quickly comprehend and implement. To improve the interpretability of risk information, multiple studies have enhanced the intuitive representation of assessment outcomes through interactive interfaces. A tower crane layout planning decision support system combining explainable deep learning models with graphical user interfaces has been proposed, enabling complex safety assessment results to be visualized and thereby reducing collision risks caused by insufficient experience or information asymmetry [48]. In addition, embedding knowledge-driven risk assessment logic into interactive systems has enabled automated coordination of construction safety risk identification, evaluation, and response [49]. To address the dispersed and implicit nature of safety requirement information in construction projects, prior studies have introduced multivariate natural language processing techniques to achieve adaptive recommendation of safety requirement categories and precise extraction of key elements, thereby enhancing the alignment between safety information and project contexts [50]. On this basis, automated safety requirement retrieval and document-level semantic association frameworks have further improved the accuracy of implicit safety requirement identification and the efficiency of information retrieval, providing more targeted support for risk-informed decision-making [51]. Furthermore, studies from communication and language interaction perspectives indicate that the way risk assessment results are expressed is itself a crucial factor influencing safety performance. Well-designed HMI language and communication strategies can significantly improve workers’ understanding and execution of collision risk control measures [52].

Under conditions of high-density equipment operation and human–machine collaborative work, the influence of HMI on construction collision risk becomes even more pronounced. Complex HMI interfaces and excessive information load may increase operators’ psychological stress, thereby inducing misoperations and collision accidents. This highlights the necessity of incorporating systematic HMI modeling approaches at the equipment design stage [53]. Further investigations have revealed that, in tower crane operation scenarios, significant differences exist in the relationship between workload and collision risk under different HMI paradigms. Specifically, human-dominated and human–machine joint interaction modes exhibit heterogeneous characteristics in the evolution of risk [54], suggesting that the role configuration within the interaction may influence the underlying risk modulation mechanisms. With the increasing application of intelligent construction equipment and robots, collision risk management under HMI conditions has gradually become a research focus. Systematic reviews indicate that HMI modules serve as a critical connection between multimodal perception, intelligent analysis, and risk response, forming the foundation for safe human–machine collaboration. Meanwhile, interaction systems based on digital twin technology have been applied to collision risk analysis in complex equipment operations. By mapping physical equipment states in real time and providing interactive monitoring interfaces, these systems effectively support remote operation and safety decision-making [55,56].

In addition, immersive and multimodal HMI technologies provide novel research avenues for construction collision risk mitigation. AR, VR, and extended reality (XR) technologies significantly enhance workers’ understanding of spatial relationships and potential collision risks by overlaying virtual risk information onto real construction environments [57,58]. Studies have shown that VR-based interactive safety training can strengthen experiences of high-risk scenarios under hazard-free conditions, thereby promoting the development of robust safety awareness [59]. At the same time, multimodal HMI integrates visual, auditory, tactile, and physiological signals to dynamically monitor workers’ cognitive load and attention states, providing support for proactive intervention in construction collision risks [60,61].

Overall, existing research generally agrees that HMI spans the perception, assessment, decision-making, and intervention stages of construction collision risk, serving as a crucial enabler for intelligent construction safety management. Human-centered HMI design effectively transforms intelligent sensing and analytical outputs into safety decisions that are understandable and actionable for construction personnel. However, current studies still tend to focus on individual technologies or specific operational scenarios, with limited consideration of the systematic integration of multisource interactive information and its long-term adaptability within complex construction environments. In the future, with the continued integration of intelligent construction technologies and artificial intelligence algorithms, the role of HMI in construction collision risk prevention is anticipated to be further expanded and refined.

3. Methodology

3.1. Experimental Design

Given that construction workers differ in body weight, height, age, educational background, and other individual characteristics, ignoring these personal factors in fatigue assessment may compromise the practical relevance of the evaluation results. Therefore, prior to the formal experiment, each participant underwent an individualized physical fatigue tolerance test, followed by a pilot experiment. The pilot experiment was required to be consistent with the formal experiment in all essential aspects. Individual maximum physical fatigue capacity was assessed using a stair-climbing task, with the highest floor level completed by each participant serving as a reference to determine their maximum physical fatigue state. Subsequently, based on this maximum fatigue level, corresponding low-fatigue and moderate-fatigue states were scientifically defined for each participant, providing personalized parameter support for subsequent fatigue classification experiments. During the fatigue induction process (stair climbing), all participants were instructed to maintain a normal pace without intentionally accelerating or decelerating, thereby minimizing the influence of unnatural movements or speed deviations and to enhance the consistency and repeatability of fatigue induction.

An intelligent tower crane operation scenario was implemented in a laboratory environment using a within-subject experimental design. Participants were exposed to three physical fatigue states—low fatigue, moderate fatigue, and high fatigue—under two experimental conditions: traditional mechanical operation and HMI. Fatigue was induced through a simulated stair-climbing task. Participants performed standardized lifting tasks in the simulated laboratory environment under each fatigue state in both experimental scenarios. During the experiment, representative eye-tracking indicators and fatigue scale measures were selected, and displacement sensors were used to measure the distance between the crane hook and the rigger. In addition, alarm devices were employed to record the number of response events under both scenarios, which served as indicators for collision risk occurrence. Through these experiments, the role of HMI in the relationship between operator fatigue state and collision risk was investigated. The experimental framework is illustrated in Figure 1.

3.2. Rationale for Experimental Scenario Design

The experimental task procedure in this study was constructed based on the general operational steps of actual tower crane lifting operations, encompassing key stages such as target positioning, hook movement control, spatial path adjustment, and personnel coordination. The spatial proportions, hook movement trajectories, and safety distance parameters in the experimental scenario were scaled and configured with reference to relevant construction codes and engineering practice data, ensuring that the logical structure of the experimental tasks remained consistent with real-world operational processes.

During the development of the experimental scenario, particular emphasis was placed on retaining the core risk-related variables directly associated with collision risk in tower crane operations. These included three-dimensional spatial distance judgment, dynamic target movement control, and visual warning feedback mechanisms. Such factors constitute the primary operational basis for risk formation in tower crane activities. By measuring and analyzing these key variables within a controlled environment, this study aims to identify the influence of fatigue states and HMI interventions on the mechanisms underlying risk formation. It should be noted that macro-level factors present on construction sites—such as organizational management variables, multi-team coordination mechanisms, and complex environmental disturbances—were not incorporated. Instead, the focus was placed on operational-level risk control mechanisms to enable a controlled examination of how HMI affects the relationship between operator fatigue and collision risk.

3.3. Physical Fatigue Induction and Classification Method

This study seeks to simulate the gradually accumulated fatigue experienced by operators during prolonged tower crane operations. Although a short-term, controllable physical workload induction method was adopted in the experiment, its theoretical basis lies in the physiological and cognitive responses following high-intensity physical activity. Specifically, such activity can increase cardiovascular load, lead to the accumulation of muscular metabolic by-products, and consume central nervous system resources, thereby resulting in decreased attention, delayed reaction time, and reduced cognitive processing capacity. Previous research has indicated that physical fatigue can influence cognitive system functioning through central fatigue mechanisms, with particularly pronounced effects in tasks requiring sustained monitoring and spatial judgment. Accordingly, physical fatigue was employed in this study as a controllable inducing variable to simulate the integrated fatigue state that operators may experience after prolonged on-site operations, with particular emphasis on its impact mechanisms on risk perception and operational control performance.

To achieve a multidimensional assessment of fatigue status, this study employed a comprehensive evaluation approach integrating subjective scales and objective physiological indicators. Subjective fatigue was measured using the FS-14 (Fatigue Scale-14). The FS-14 is a brief and effective fatigue assessment instrument used to measure individuals’ perceived fatigue and the impact of fatigue on daily functioning [62]. The FS-14 was jointly developed in 1992 by Professor Chalder from the Department of Psychological Medicine at King’s College Hospital and Dr. G. Berelowiyz from the Royal London Hospital (Queen Mary University) [63]. Participants were required to carefully read each item and select either “Yes” or “No” based on the option that best reflected their current condition. The physical fatigue score was obtained by summing the scores of the first eight items (Items 1–8), while the mental fatigue score was calculated by summing the scores of the remaining six items (Items 9–14). The total fatigue score was computed as the sum of the physical and mental fatigue scores. The maximum possible scores are 8 for physical fatigue, 6 for mental fatigue, and 14 for total fatigue, with higher scores indicating more severe fatigue [64].

Given that tower crane operation tasks are inherently characterized by high visual dependence and sustained attentional demands, reliance solely on subjective scales is insufficient to comprehensively capture cognitive fatigue variations. Therefore, this study further employed a Pupil Core eye-tracking system to collect objective eye-movement data, including blink frequency, fixation duration, saccade frequency, and pupil diameter variation. Previous studies have demonstrated that fluctuations in pupil diameter are closely associated with cognitive load, blink frequency is significantly correlated with sustained attention capacity, and prolonged fixation duration typically reflects reduced information-processing efficiency. Accordingly, eye-movement parameters were utilized to characterize cognitive fatigue manifestations potentially induced by physical workload.

Fatigue classification was determined using an “individual baseline-relative change” approach. First, each participant’s eye-movement indicators under resting conditions were recorded as individual baseline values. Following fatigue induction, both eye-movement metrics and FS-14 scores were reassessed. Fatigue levels were categorized according to the following criteria. Low fatigue was defined as a slight increase in FS-14 scores relative to baseline, accompanied by minor changes in eye-movement indicators, typically manifested as a slight increase in blink frequency, relatively stable pupil diameter, and largely maintained cognitive processing capacity. Moderate fatigue was identified when both physical and mental dimension scores of the FS-14 were significantly higher than baseline values, and eye-movement indicators exhibited noticeable changes, including increased blink frequency and fixation duration, as well as enhanced pupil diameter fluctuations, suggesting reduced efficiency in attentional resource allocation. High fatigue was defined when the total FS-14 score approached the upper bound of the individual’s maximum fatigue testing range, and eye-movement indicators displayed pronounced abnormal fluctuations, including a marked increase in blink frequency, reduced regularity of saccadic movements, and an overall decreasing trend in pupil diameter, reflecting substantial impairment in cognitive control capability. The core assumption underlying this classification method is that when physical workload exceeds a certain threshold, central nervous system regulatory mechanisms influence cognitive system functioning. Therefore, fatigue classification was not based solely on physical exertion levels but was determined through an integrated assessment combining physical fatigue, mental fatigue dimensions, and objective cognitive indicators. Cross-validation across multiple measures enhances the theoretical robustness and measurement reliability of the fatigue categorization.

In addition, to ensure participant safety, heart rate was continuously monitored during the experiment. Participants wore heart rate belts, and individual maximum heart rate thresholds were calculated using the Gellish formula (Maximum Heart Rate = 206.9 − 0.67 × age). The experiment was immediately terminated when heart rate approached the safety threshold to prevent physiological risks associated with excessive workload. Through these procedures, the study established, under controlled experimental conditions, an associative pathway linking physical workload, cognitive performance, and risk-related behavior. This approach strengthens the theoretical connection between the fatigue induction protocol and fatigue states observed in real tower crane operations, while clarifying the criteria and logical basis for fatigue classification.

3.4. Collision Risk Indicators

Based on previous studies, operational data recorded under two experimental scenarios and different fatigue states were selected as the basis for analyzing the effects of HMI on collision risk. Displacement sensors were used to record the safety distance between the crane hook and the rigger, which serves as a key indicator for evaluating whether the intelligent tower crane maintains a safe separation from the rigger during cooperative operations [65], thereby quantifying potential collision risk during crane operation. A smaller safety distance indicates that the crane trajectory is closer to the rigger, significantly increasing the likelihood of collision and reflecting reduced risk control capability. By comparing this indicator across different fatigue states under human–machine interaction and traditional mechanical operation scenarios, the role of HMI in supporting operators’ risk perception and fine-grained operational control can be effectively identified, allowing for a scientific evaluation of its effectiveness in reducing collision risk under fatigue conditions. The number of alarm responses was used as an additional indicator to measure the frequency of collision risk events, providing a direct reflection of how often collision risks occurred under different fatigue states [66]. A higher number of alarm responses may indicate that operators exhibited more hazardous behaviors while controlling the tower crane or that they experienced insufficient environmental perception or delayed reactions, particularly under high-fatigue conditions. This indicator, combined with the system’s warning mechanism, enables effective assessment of the role of HMI in assisting operators with risk avoidance, while also reflecting operators’ adaptability to alarm feedback. Therefore, this study developed a two-dimensional quantitative framework for collision risk assessment based on “spatial distance-warning frequency.” Within this framework, the spatial distance indicator represents the intensity of risk exposure, while warning frequency reflects the tendency of risk activation. The two dimensions complement each other, thereby enhancing the structural integrity and interpretive consistency of the risk assessment model. As both indicators are closely related to collision risk and can effectively capture the influence of human–machine interaction on the relationship between operator fatigue and collision risk, they were selected as the key analytical metrics in this study.

According to research in ergonomics and construction safety, the recommended safe operating radius for personnel within the lifting area of actual tower crane operations—considering dynamic buffer distance—is generally not less than 2 m, so as to reduce collision risks caused by load swing and operational errors. In this study, the experimental platform was constructed using a 1:10 scaled model. To ensure consistency in spatial judgment, the safety radius was proportionally set at 200 mm. Based on this safety buffer radius, collision risk was classified from the perspective of risk exposure level. When the operational distance exceeded 200 mm, the target was considered to be outside the safety buffer zone and was categorized as “no risk.” When the distance ranged between 100 and 200 mm, the target was deemed to have entered the safety buffer zone but not yet approached the critical collision boundary, and was therefore classified as “low risk.” When the distance was less than 100 mm, the safety boundary was considered to be significantly encroached upon, representing a high-exposure interval and categorized as “high risk.” It should be emphasized that the risk classification adopted in this study is operational in nature, intended to characterize behavioral variation trends across different risk intervals, rather than to directly predict the probability of actual construction accidents.

3.5. Data Analysis

Statistical analyses were conducted on the experimental data to investigate the relationship between operators’ physical fatigue levels and collision risk during tower crane lifting operations, and to evaluate the mitigating effects of the HMI operation mode on collision risk under fatigue conditions. Physical fatigue level was treated as the independent variable and classified into three categories—low fatigue, moderate fatigue, and high fatigue—based on the fatigue test results. All participants performed lifting tasks under each fatigue condition, resulting in a within-subject repeated-measures data structure.

Given that traditional mechanical operation and HMI operation constituted two independent experimental scenarios, statistical analyses were performed separately for each scenario. Within each scenario, one-way repeated-measures analysis of variance (RM-ANOVA) was first applied to compare collision risk performance across different fatigue states, to examine the overall effect of physical fatigue level on lifting operation safety. The significance level was set at 0.05 (p < 0.05), and all statistical analyses were conducted using SPSS software, (version 27.0, IBM Corp., Armonk, NY, USA). Prior to performing RM-ANOVA, the assumptions of the method were examined. Although each participant was measured multiple times under different fatigue states, measurements across different participants were independent, satisfying the independence assumption, while measurements within the same participant were correlated, making the repeated-measures design appropriate. The Shapiro–Wilk test was used to assess the normality of the dependent variables under each fatigue condition, with p > 0.05 indicating approximate normality. In addition, Mauchly’s test of sphericity was conducted to examine the homogeneity of variances of the differences among repeated measures. When the sphericity assumption was violated (Mauchly’s test p < 0.05), the Greenhouse–Geisser correction was applied to adjust the degrees of freedom. To quantify the magnitude of the effects, partial η² was reported to assess the practical impact of fatigue levels on collision risk.

On this basis, to further quantitatively analyze the effects of physical fatigue level on collision risk in tower crane operations and to enhance the robustness of statistical inference while controlling for individual differences, linear mixed-effects models (LMMs) were constructed separately for the traditional mechanical operation scenario and the HMI scenario. The safety distance between the crane hook and the rigger was specified as the dependent variable, while physical fatigue level (low, moderate, and high) was included as a fixed effect to evaluate the impact of different fatigue states on safety distance. Considering that each participant took part in multiple trials under different fatigue conditions and that repeated measurements were therefore not independent, participant ID was included as a random effect to account for inter-individual variability and to control for the repeated-measures structure. The model parameters were estimated using the Restricted Maximum Likelihood (REML) method. The F-values, p-values, and effect sizes (Cohen’s d or partial η²) of the fixed effects were reported to evaluate the magnitude of the influence of physical fatigue on safety distance. The significance of the fixed-effect terms was examined to assess the magnitude of the effect of physical fatigue level on safety distance. The significance level for all statistical tests was set at 0.05. Model fitting and parameter estimation were performed using the Linear Mixed Models module in SPSS software (Version 27). By establishing LMMs separately for the two experimental scenarios and comparing the estimated fixed-effect parameters, the mitigating effects of the HMI operation mode on collision risk under different fatigue states were revealed.

In addition, collision alarm frequencies under different fatigue states were incorporated to provide a supplementary analysis of collision risk levels in both traditional mechanical operation and HMI operation scenarios, thereby offering a comprehensive reflection of operators’ safety performance under varying fatigue conditions. By comparing the trends of fatigue effects between the traditional mechanical operation and HMI operation scenarios, the effectiveness of the HMI system in mitigating collision risk during tower crane lifting operations under operator fatigue was systematically evaluated.

4. Experiment

4.1. Participants

Participants were recruited through collaboration with construction enterprises. Prior to the experiment, an a priori power analysis was conducted using G*Power 3.1.9.7 software (version 3.1.9.7, University of Düsseldorf, Dusseldorf, Germany), with a one-way repeated-measures analysis of variance as the planned study design. Specifically, a moderate effect size (f = 0.25), a significance level of α = 0.05, and a statistical power (1 − β) of 0.80 were specified. This preliminary power analysis provided an effective means of controlling statistical power and guiding sample size determination before initiating the actual study. The sample size was determined based on the research design, variable dimensionality, and statistical power considerations, and it meets the standards of mainstream empirical research. The power analysis indicated that a minimum of 28 participants was required to detect the expected effect. Accordingly, a total of 28 healthy adult construction workers were recruited after screening, including 17 males and 11 females. Participants were aged between 35 and 45 years and had undergone medical examinations at a Grade III, Class A hospital, confirming the absence of neurological disorders, musculoskeletal diseases, or other health conditions that could affect experimental performance. All participants had more than five years of professional experience in formal construction enterprises, and their safety training assessment results over the past two years were rated as excellent. All participants provided written informed consent and received detailed explanations of the experimental procedures and safety instructions prior to participation. After completing the experiment, participants received financial compensation. Participants were instructed to maintain adequate sleep on the day before the experiment, avoid staying up late or excessive alcohol consumption, refrain from taking medications, avoid strenuous physical activity, and abstain from spicy or greasy foods, to minimize potential biases caused by physical discomfort on the experimental day. In addition, each participant was assigned an experimental identification number from “1” to “28”, and the experimental sessions were conducted sequentially according to this numbering.

4.2. Experimental Scenarios

Two simulated operation scenarios based on an intelligent tower crane system were established in this experiment: a traditional mechanical operation scenario and an HMI operation scenario. Each scenario consisted of an operator control area and a lifting operation area to support the investigation of crane hoisting tasks. As shown in Figure 2 and Figure 3, the overall layout of the experimental setup includes the main structure of the tower crane, the hook movement area, the operator control zone, and the lifting operation area. The experimental site was configured with three typical lifting zones. Operators were required to use the intelligent tower crane interaction system to sequentially control the crane hook, lift target objects from the starting point, avoid dynamic obstacles (intelligent mobile robots), and accurately place the objects into the designated target zones. The figure clearly illustrates the hook trajectory, the activity range of the dynamic obstacles, and the spatial relationships between personnel and equipment, thereby providing an intuitive representation of the structural layout of the experimental environment.

A small-scale, high-precision intelligent tower crane equipped with a real-time monitoring and intelligent feedback system was employed in the experiment, enabling operators to perform complex lifting operations with system assistance. During operation, participants monitored the crane’s operational status and target positions via a display interface at the control console and executed lifting tasks using control devices. In the HMI operation scenario, operators were required to make timely judgments and adaptive adjustments based on system feedback and alarm responses to ensure operational safety and enhance HMI efficiency. In contrast, in the traditional mechanical operation scenario, the HMI system of the intelligent tower crane was deactivated, and operators relied solely on their own observation of the working environment and manual decision-making to complete the lifting tasks, thereby realistically simulating the actual operating mode of conventional tower cranes.

In addition, as shown in Figure 4, the safety distance in this study was defined as the Euclidean distance between the geometric center of the crane hook and the reference position of the signal worker. By continuously collecting the spatial coordinate data of both entities, the dynamic variation in this distance was calculated to quantify the collision risk under different fatigue states. This measurement approach objectively reflects the safety clearance between personnel and equipment, thereby improving the accuracy and reproducibility of collision risk assessment.

4.3. Experimental Procedure

Each participant completed experiments under the same fatigue state across two experimental scenarios within a single day. To control for potential time-related effects, all participants performed the experiments at a consistent time each day. After completing the pilot test, participants advanced to the formal experiments. First, under the traditional mechanical operation scenario, participants performed a stair-climbing task to induce low, moderate, and high levels of fatigue, respectively. They then completed the subjective fatigue questionnaire and wore the eye-tracking device, which was adjusted and calibrated before entering the preliminary experimental stage. Eye-tracking indicators were analyzed to verify the attainment of target fatigue states and to minimize potential subjective biases. If the corresponding fatigue level was confirmed, participants were allowed to proceed to the formal experiment; otherwise, fatigue induction was continued until the required fatigue state was achieved. After reaching the designated fatigue level, participants completed three standardized lifting tasks, with a displacement sensor continuously recorded the safety distance between the crane hook and the rigger.

Under the HMI operation scenario, the experimental procedures were largely consistent with those of the traditional mechanical operation scenario. During this process, operators monitored the real-time operational status of the tower crane through the intelligent crane operation platform and made adjustments based on system feedback. The displacement sensor continuously measured the safety distance between the crane hook and the rigger, while the alarm system responded to sensor data in real time. When the crane hook approached the rigger beyond a safe threshold or when a potential collision risk was detected, the alarm system was triggered to warn the operator of improper actions. The operator then adjusted the operation according to maintain the lifting equipment within a safe range and prevent collisions. The intelligent tower crane operating system responded in real time to the operator’s performance. Specifically, warnings were issued when the load was excessive or when the hook descended below a safe level, and emergency braking measures were activated under critical conditions to ensure the safety of the human–machine collaborative process. By comparatively analyzing the safety distances maintained by operators under different fatigue states in HMI and traditional mechanical operation scenarios, this study aimed to elucidate the role of HMI in modulating the relationship between operator fatigue and collision risk, thereby providing theoretical support for enhancing operational safety and HMI efficiency in intelligent tower crane operations. A detailed operational flowchart is shown in Figure 5.

5. Experimental Results

5.1. Descriptive Statistics

In both the traditional mechanical operation and human–machine interaction operation scenarios, the minimum safety distance between the crane hook and the rigger, as well as the number of collision alarm responses, were recorded and analyzed under different operator fatigue states (low, moderate, and high fatigue).

As shown in

Table 1. Safety Distance for Each Fatigue Condition in Traditional Mechanical Operations.

Metric	F1		F2		F3
	Mean	SD	Mean	SD	Mean	SD
Distance	228.18	27.19	149.15	16.54	96.54	8.53
Number of Alarms	0.71	0.99	2.96	0.19	3	3

and

Table 2. Safety Distance for Human–Machine Interaction Operations Under Various Fatigue Conditions.

Metric	F1		F2		F3
	Mean	SD	Mean	SD	Mean	SD
Distance	317.46	27.33	243.17	10.61	170.23	17.41
Number of Alarms	0	0	0.18	0.39	2.61	0.50

Note: F1 = Low Fatigue; F2 = moderate Fatigue; F3 = High Fatigue.

, in the traditional mechanical operation scenario, the mean minimum safety distances maintained by operators under low, moderate, and high-fatigue states were 228.18 mm, 149.15 mm, and 96.54 mm, respectively, while the mean numbers of alarm responses were 0.71, 2.96, and 3. In the HMI scenario, the corresponding mean minimum safety distances were 317.46 mm, 243.17 mm, and 170.23 mm, and the mean alarm responses were 0, 0.18, and 2.61, respectively. These results demonstrate that, across both operational scenarios, increasing fatigue levels were associated with a progressive reduction in safety distance and a corresponding increase in the frequency of collision alarms. Specifically, under high-fatigue conditions, the minimum safety distance reached its lowest level, while the frequency of collision alarms peaked, indicating a substantially elevated risk of collision. In comparison, in the HMI scenario, the safety distances under all fatigue states were generally greater than those in the traditional mechanical operation scenario, and the number of alarms was significantly reduced. This suggests that the human–machine interaction system can effectively improve operational safety under fatigued conditions.

As illustrated in Figure 6a, within the traditional mechanical operation scenario, operator performance varied markedly across different fatigue states, resulting in significant differences in collision risk. This indicates a significant correlation between fatigue state and collision risk. Overall, under low-fatigue conditions, most participants maintained operational distances ranging from 170 to 280 mm, whereas under moderate fatigue this range decreased to 120–180 mm, and under high-fatigue conditions the majority of participants maintained distances of only 80–120 mm. This demonstrates that collision risk gradually increases with rising fatigue levels. Specifically, under low-fatigue conditions, most operations were free of collision risk, with only a few instances classified as low risk. Moderate-fatigue conditions corresponded primarily to low risk levels, whereas high-fatigue conditions were significantly associated with high risk. Although some instances remained in the low-risk category, the overall risk level clearly increased. Figure 6b further confirms this trend. Distance data recorded by the displacement sensors for each lifting task consistently decreased as fatigue intensified, indicating a continuous increase in collision risk during operation. This pattern suggests a clear correlation between operator fatigue state and collision risk.

The line chart results in the HMI scenario, shown in Figure 7a, indicate a certain correlation between operator fatigue state and collision risk. Specifically, under low-fatigue conditions, most participants maintained safety distances of 270–370 mm during operation, whereas under moderate fatigue this distance decreased to 225–260 mm, and under high-fatigue conditions it further declined to a range of 140–200 mm. Regarding collision risk, no significant risk was observed under low or moderate-fatigue conditions, whereas under high-fatigue conditions, although no actual collisions occurred, the majority of operations were categorized as posing a low level of risk. Figure 7b further corroborates these findings and reveals additional insights. Data collected by displacement sensors throughout the entire lifting task indicate that as fatigue increased, the safety distance between the hook and the rigger steadily decreased, reflecting the effect of accumulated fatigue on the elevation of collision risk.

5.2. Repeated Measures ANOVA Results

5.2.1. Traditional Mechanical Operation Scenario

In the traditional mechanical operation scenario, a one-way repeated measures ANOVA was conducted on the minimum safety distance between the hook and the rigger, with operator fatigue level as the within-subject factor. The measured results are recorded in

Table 3. Results of One-Way Repeated Measures ANOVA.

Source	df	F-Value	p-Value	Partial η2
Fatigue (traditional mechanical operation)	2, 54	362.125	<0.001	0.931
Fatigue (human–machine interaction operation)	2, 54	441.403	<0.001	0.942

. The results indicated that fatigue level had a highly significant effect on safety distance (362.125, p < 0.001). These results indicate that variations in fatigue state significantly explain differences in maintained safety distance, with a very large effect size, suggesting that fatigue may be an important factor influencing operational safety under traditional mechanical operation conditions.

Moreover,

Table 4. Pairwise Comparisons of Safety Distances Between Fatigue Levels in Traditional Mechanical Operations.

Pairwise Comparison	Fatigued State	p-Value
Low fatigue	Moderate fatigue	<0.001
	High fatigue	<0.001
Moderate fatigue	High fatigue	<0.001

records the comparative results of safety distances between different fatigue levels in traditional mechanical operations. Post hoc comparisons revealed that the minimum safety distance under low fatigue was significantly greater than that under moderate and high fatigue (p < 0.001), and the minimum safety distance under moderate fatigue was also significantly greater than that under high fatigue (p < 0.001). These results indicate that increasing operator fatigue under traditional mechanical operation conditions substantially elevates the potential collision risk during lifting tasks.

5.2.2. HMI Operation Scenario

In the HMI operation scenario, a one-way repeated measures ANOVA was similarly conducted to assess the effect of fatigue level on safety distance, and the results are recorded in

Table 5. Pairwise Comparisons of Safety Distances Across Fatigue Levels in HMI Operations.

Pairwise Comparison	Fatigued State	p-Value
Low fatigue	Moderate fatigue	<0.001
	High fatigue	<0.001
Moderate fatigue	High fatigue	<0.001

. The results indicated that operator fatigue continued to exert a significant effect on safety distance (441.403, p < 0.001), although the extent of variation in safety distance across different fatigue states was substantially smaller than that observed in the traditional mechanical operation scenario. Post hoc comparisons indicated that the minimum safety distance under low fatigue was significantly greater than that under moderate and high fatigue (p < 0.001), and the difference between moderate and high fatigue was also statistically significant (p < 0.001). Nevertheless, the descriptive statistics indicate that the range of safety distance variation across fatigue levels in the HMI operation scenario was noticeably smaller than in the traditional mechanical operation scenario, suggesting that the HMI system partially attenuates the detrimental effects of increasing fatigue on safety distance, thereby enhancing operational stability and safety margins.

In summary, the results of the one-way repeated measures ANOVA consistently indicate that, in both operation scenarios, operator fatigue levels have a significant effect on the minimum safety distance maintained during lifting operations. However, compared with the traditional manual operation scenario, the HMI mode generally maintained larger safety distances across different fatigue states and demonstrated a mitigating trend in collision risk under fatigue conditions.

5.3. Linear Mixed-Effects Model Analysis Results

To evaluate the effects of different fatigue levels on the measured indicators, a linear mixed-effects model (LMM) was constructed in this study. Fatigue level was treated as a fixed effect, with three categories (low, moderate, and high), using the high-fatigue group as the reference, while random effects were incorporated to control for inter-individual variability.

5.3.1. Linear Mixed-Effects Model Results for the Traditional Mechanical Operation Scenario

Next,

Table 6. Linear Mixed Effects Model Results for Traditional Mechanical Operation Scenarios.

Effect	β	SE	T-Value	p-Value	95% Confidence Interval
Intercept	96.54	4.02	24.02	<0.001	[88.66, 104.41]
Fatigue (Low vs. High)	131.64	3.1	42.4	<0.001	[125.54, 137.74]
Fatigue (Moderate vs. High)	52.62	3.1	16.95	<0.001	[46.52, 58.72]

records the results of the linear mixed effects model (traditional Mechanical Operation Scenarios). Under traditional mechanical operation conditions, the model results showed an intercept of 96.54 (standard error, SE = 4.02, t = 24.02, p < 0.001), indicating that the predicted mean value of the response variable under high fatigue was approximately 96.54. Compared with the high-fatigue group, the effect estimate for the low-fatigue group was 131.64 (SE = 3.10, t = 42.40, p < 0.001, 95% CI: 125.54–137.74), and the effect estimate for the moderate-fatigue group was 52.62 (SE = 3.10, t = 16.95, p < 0.001, 95% CI: 46.52–58.72), both significantly higher than the high-fatigue group. These results indicate that, under the traditional mechanical operation mode, fatigue level exerts a significant effect on the measured indicator, with the response variable exhibiting a pronounced increasing trend as fatigue decreases.

5.3.2. Linear Mixed-Effects Model Results for the HMI Operation Scenario

Besides,

Table 7. Linear Mixed Effects Model Results for HMI operation scenario.

Effect	β	SE	T-Value	p-Value	95% Confidence Interval
Intercept	170.23	4.48	38.01	<0.001	[161.45, 179.01]
Fatigue (Low vs. High)	147.24	3.57	41.25	<0.001	[140.22, 154.25]
Fatigue (Moderate vs. High)	72.94	3.57	20.43	<0.001	[65.93, 79.95]

records the results of the linear mixed effects model (HMI operation scenario). Under HMI operation scenario, the model results showed an intercept of 170.23 (SE = 4.48, t = 38.01, p < 0.001), representing the baseline level under high-fatigue conditions. Compared with the high-fatigue group, the effect estimate for the low-fatigue group was 147.24 (SE = 3.57, t = 41.25, p < 0.001, 95% CI: 140.22–154.25), and the effect estimate for the moderate-fatigue group was 72.94 (SE = 3.57, t = 20.43, p < 0.001, 95% CI: 65.93–79.95), both reaching statistical significance. These results also indicate that as fatigue level decreases, the response variable increases significantly, suggesting that fatigue status has a stable and significant positive effect on the indicator.

Overall, the analysis of both operation modes indicates that fatigue level is a significant predictor of the measured indicators. Under both traditional mechanical operation scenario and HMI operation scenario, the response values of the low- and moderate-fatigue groups were significantly higher than those of the high-fatigue group (p < 0.001), showing a clear dose–response relationship. This finding underscores the importance of accounting for operator fatigue in construction-related work environments and operational management to optimize operational practices, improve working conditions, and enhance overall safety and efficiency.

5.4. Collision Risk Analysis Under Traditional Mechanical Operation Scenarios and HMI Operation Scenarios

5.4.1. Safety Distance

To further assess the impact of the HMI operation scenario on collision risk under different fatigue conditions, this study selected two representative fatigue states, moderate fatigue and high fatigue, for a comparative analysis between traditional mechanical operation scenarios and HMI operation scenarios. Figure 8 illustrates the distance changes under moderate-fatigue conditions before and after interaction. In the traditional mechanical operation scenario, the distance between the hook and the rigger was primarily distributed within the 120–180 mm range, indicating a certain collision risk. In contrast, within the HMI operation scenarios, the safety distance increased significantly to 225–260 mm, indicating enhanced operational performance and a reduction in risk level from “low risk” to “no risk”. This indicates that HMI system can effectively compensate for the decline in operator control ability caused by fatigue, thereby significantly reducing interaction risk. Figure 9 presents the experimental results under high-fatigue conditions. In the traditional mechanical operation scenario, the average safety distance remained between 80 and 120 mm, corresponding to a high collision risk level and indicating substantial impairments in both operational stability and safety. However, in the HMI operation scenario, the average distance increased to 120–180 mm, reducing the risk level from “high risk” to “low risk”. Although the safety distance still did not reach the non-fatigue standard, overall safety was significantly improved. It can thus be inferred that, under the experimental conditions of this study, HMI exhibited a certain mitigating trend with respect to collision risk indicators under high-fatigue states; however, its practical effectiveness requires further validation.

Based on the comparative results, it can be concluded that the HMI system elevates the operational status from “low risk” to “no risk” under moderate-fatigue conditions and reduces the “high risk” level to “low risk” under high-fatigue conditions, thereby significantly enhancing overall operational safety and precision. This change indicates that HMI can actively intervene in the operational process through assisted control, intelligent correction, and information feedback, effectively counteracting the decline in operational performance and the increase in collision risk caused by fatigue. The HMI mode may therefore be an important factor in promoting the safe operation of intelligent tower cranes during high-risk tasks. Analysis of key indicators before and after interaction suggests that this technology may have positive effects on improving operational accuracy, increasing safety distances, and reducing collision risks. Therefore, in future high-intensity work environments, the introduction and optimization of HMI will be a key pathway to enhancing the operational safety of intelligent equipment.

5.4.2. Alarm Responses

To examine differences in collision risk across operational scenarios, this study analyzed alarm response counts under both the traditional mechanical operation scenario and the HMI operation scenario (The test results are recorded in

Table 8. Number of alarm responses in two scenarios.

Response Times
Number	TM	HMI	TM	HMI	TM	HMI
1	0	0	3	0	3	3
2	0	0	3	0	3	2
3	0	0	3	0	3	3
4	1	0	3	1	3	2
5	1	0	2	0	3	2
6	3	0	3	1	3	3
7	3	0	3	0	3	3
8	0	0	3	0	3	3
9	2	0	3	0	3	3
10	0	0	3	0	3	2
11	0	0	3	0	3	3
12	0	0	3	0	3	2
13	1	0	3	0	3	3
14	0	0	3	1	3	2
15	0	0	3	0	3	2
16	1	0	3	0	3	2
17	0	0	3	0	3	2
18	0	0	3	0	3	3
19	1	0	3	1	3	3
20	0	0	3	0	3	3
21	3	0	3	1	3	3
22	2	0	3	0	3	3
23	0	0	3	0	3	3
24	1	0	3	0	3	3
25	0	0	3	0	3	2
26	1	0	3	0	3	3
27	0	0	3	0	3	2
28	0	0	3	0	3	3
Sum	20	0	83	3	84	73

Note: TM = Traditional Mechanical; HMI = Human–Machine Interaction.

). The results reveal a marked difference in alarm frequency between the two scenarios, with fatigue level exerting a significant influence.

In the traditional mechanical operation scenario, the overall alarm response counts were comparatively high, with 20, 83, and 84 occurrences observed under low, moderate, and high-fatigue conditions, respectively. This further confirms that collision risk increases with rising fatigue levels. Under high-fatigue conditions, the alarm trigger frequency was significantly higher than under moderate- and low-fatigue conditions, suggesting that accumulated fatigue in traditional mechanical operations may substantially increase the likelihood of operational errors or abnormal events, leading to more frequent alarms. In contrast, under low-fatigue conditions, the number of alarms remained low, indicating relatively stable operation.

In the HMI operation scenario, the overall alarm response counts were substantially lower, with 0, 3, and 73 occurrences observed under low-, moderate-, and high-fatigue conditions, respectively, indicating that alarm frequency increases markedly with rising fatigue. When operators were in low- and moderate-fatigue states, the HMI system remained stable, and no notable alarms were triggered, reflecting good stability and fault tolerance. However, under high-fatigue conditions, system instability began to emerge, and alarms were triggered more frequently. This suggests that high fatigue impairs perceptual and control accuracy for operators, thereby compromising collaborative safety between the intelligent tower crane and riggers and substantially increasing the risk of collision. Comparisons show that under low- and moderate-fatigue conditions, alarm counts in the HMI operation scenario were markedly lower than in the traditional mechanical operation scenario. Even under high-fatigue conditions, alarm frequency remained below that of the traditional mechanical scenario, indicating that the HMI system can mitigate the negative effects of fatigue on operational safety to some extent.

Overall, the results from both operation scenarios indicate that the operation mode significantly affects alarm response counts. Compared with traditional mechanical operation, HMI operation scenario consistently exhibited lower alarm frequencies across all fatigue levels, with the advantage being particularly pronounced under low- and high-fatigue conditions. The above results indicate that, under experimental conditions, HMI exhibited a tendency to reduce the likelihood of collision events and may contribute to enhancing the safety and operational reliability of the system to a certain extent. The results provide strong evidence for optimizing operational modes in construction-related environments and highlight the importance of integrating HMI systems under high workload or fatigue-prone conditions.

6. Discussion

In high-risk construction environments, previous studies suggest that physical fatigue may not only impair workers’ physical performance but may also indirectly influence operational safety by affecting perception, judgment, and motor control. Under controlled laboratory conditions, this study systematically examined the relationship between operators’ physical fatigue and collision risk across different operational scenarios and explored the moderating effect of HMI under fatigue. The findings observed in the experimental context reveal certain tendencies and may provide empirical support for understanding the fatigue-risk mechanism and the potential role of intelligent assistance systems in construction safety.

6.1. Relationship Between Physical Fatigue and Collision Risk

Under the laboratory-simulated traditional mechanical operation scenario, as physical fatigue gradually increased, the minimum safety distance between the hook and the riggers significantly decreased, and the alarm response count increased markedly, resulting in a continuous rise in overall collision risk. The same pattern was observed in the HMI operation scenario. These findings suggest that, within the experimental setting, physical fatigue may weaken operators’ control capability and risk avoidance performance in complex tasks. The underlying mechanism is likely related to fatigue-induced decreases in alertness, reaction speed, and fine motor control, which impair perception and judgment of critical spatial information during lifting operations. As fatigue levels increase, operators are more prone to delayed responses, unstable motion amplitude control, and insufficient risk anticipation, leading to deviations of the hook trajectory from safe zones. These findings align with previous research showing that physical fatigue significantly diminishes situational awareness and hazard recognition, thereby negatively affecting safety performance. Empirical evidence indicates that fatigue reduces construction workers’ hazard recognition and risk assessment capabilities, reflecting a decrease in attentional resources and deterioration of safety performance [67]. Moreover, the reduction in safety distance implies a smaller tolerance margin for potential hazards under fatigue, meaning that operational errors are more likely to trigger dangerous events. This partially explains why collision risk in traditional operation scenarios rapidly escalates to high-risk levels under high-fatigue conditions. It should be emphasized that these findings are derived from a laboratory-based simulation and reflect associations between fatigue levels and risk indicators rather than direct predictions of real-world construction accident rates.

6.2. Mitigating Effect of HMI on Fatigue-Induced Collision Risk

Compared with traditional mechanical operations, under the controlled laboratory conditions established in this study, the HMI operation scenario exhibited distinct risk variation patterns. The results indicate that, at the same fatigue level, the safety distance between the hook and the riggers in the HMI operation scenario was generally greater than in the traditional scenario, with a significant reduction in alarm counts. This difference was particularly pronounced under moderate- and high-fatigue conditions. Specifically, within the experimental setting, collision risk under moderate fatigue in the HMI scenario showed a tendency to shift from low-risk toward no-risk levels, while under high fatigue, risk levels exhibited a decreasing trend from high-risk toward lower-risk categories. These findings suggest that, in the present laboratory context, the HMI system may provide a certain degree of safety compensation when operators experience fatigue.

The underlying mechanism may be that HMI systems reduce the operator’s reliance on sustained high-intensity attention and precise spatial judgment through real-time risk alerts, distance monitoring, and alarm feedback. When operators experience fatigue-induced declines in perception and decision-making, the system’s auxiliary information compensates for cognitive resource limitations, guiding timely adjustment of operational behavior and preventing proximity to hazards or collisions. Thus, under the laboratory conditions of this study, the regulatory effect of HMI appears to show a tendency to mitigate fatigue-related increases in collision risk, although the magnitude of this effect requires further verification.

Furthermore, even under high fatigue, operators’ control stability over lifting operations remained superior in the HMI operation scenario compared with traditional mechanical operations. This tendency suggests that, in the laboratory environment, HMI may not only reduce the probability of collision events but also contribute to improved operational controllability to some extent. From a human factors perspective, the HMI system may reduce cognitive and perceptual workload under fatigue by reallocating part of the information-processing demands. This “cognitive offloading” effect allows operators to focus limited attentional resources on critical operational decisions rather than continuously monitoring complex environmental information, thereby improving overall safety performance. This observation is generally consistent with prior studies on intelligent assistance systems under high workload or fatigue conditions and may provide reference for the application of HMI in complex construction equipment operations, although its broader applicability should be further examined in real-world construction settings.

6.3. Mechanistic Transferability Analysis

Although this study was conducted using a laboratory-based intelligent tower crane operation platform under controlled experimental conditions, the core variables of interest are not specific to any particular construction process. Rather, they pertain to the effects of fatigue on cognitive processing capacity, reaction time, and spatial distance judgment accuracy. These abilities reflect human cognitive mechanisms that are relatively stable across contexts, and their patterns of change generally do not fundamentally depend on the specific operational scenario. Extensive research in ergonomics has shown that fatigue systematically reduces the efficiency of attentional resource allocation, prolongs reaction times, and impairs spatial judgment accuracy, thereby affecting individuals’ ability to manage risks in dynamic environments.

In this study, the observed trends—namely, decreased ability to maintain safety distance and increased exposure to collision risk with rising fatigue levels—essentially reflect the manifestation of cognitive function degradation in three-dimensional operational tasks. Therefore, at the mechanistic level, the relationship between fatigue and spatial risk control capacity exhibits a degree of theoretical stability, suggesting that these trends may be transferable across different contexts. Similarly, the role of the HMI system in this study primarily operates through visual cues, distance feedback, and risk alarms that enhance the salience of hazards, thereby triggering attention capture and decision-correction processes. Such information-presentation mechanisms are not specific to tower crane operations but rely on general principles of human information processing and attentional regulation. Consequently, the modulatory effects of HMI on risk perception and operational control also possess a degree of cross-task consistency at the theoretical level.

It should be emphasized that the term “transferability” in this study refers primarily to mechanistic insights, rather than the direct generalization of numerical thresholds or risk proportions. Variations in risk intensity, organizational pressure, and coordination complexity may exist across different construction environments, and practical application still requires further validation under actual field conditions.

6.4. Practical Implications

The findings of this study offer new technical pathways for improving tower crane operation safety. It was observed that under traditional mechanical operations, as operator fatigue increased, the safety distance between the hook and the riggers decreased significantly, and collision risk increased markedly. Conversely, under HMI operation scenario, collision risk could be effectively controlled even when operators were under moderate- or high-fatigue conditions. This indicates that HMI systems can provide important safety compensation under operator fatigue, effectively mitigating the decline in control ability and the increase in risk exposure caused by fatigue.

Therefore, in high-intensity, long-duration, or high-risk lifting operations, prioritizing the use of intelligent tower crane systems with HMI functions can reduce the risk of collisions caused by operator fatigue. Particularly in construction environments where fatigue cannot be completely avoided, HMI systems serve as an effective engineering control measure, compensating for the safety hazards arising from diminished perception, judgment, and operational precision. Moreover, by quantifying collision risk through safety distance and alarm counts, this study provides actionable metrics for on-site risk monitoring and safety management. Construction managers can use operator fatigue levels and system alarm information to dynamically assess operational risk, allowing timely adjustment of work pace or targeted interventions.

The results of this study further emphasize the importance of systematic management of operator fatigue. Although physical fatigue cannot be fully eliminated during construction, reasonable work duration scheduling, optimized shift arrangements, and integration with HMI assistance can mitigate its adverse effects on operational safety. Particularly during critical high-risk work phases, the introduction of intelligent assistance and risk alert mechanisms can help ensure overall site safety.

7. Conclusions

This study employed a laboratory-based tower crane simulation platform to conduct comparative experiments between conventional mechanical operation and HMI operation, aiming to investigate the relationship between operator physical fatigue and collision risk, as well as the potential moderating effect of HMI under different fatigue levels. The conclusions drawn from the experimental results should be interpreted within the context of the controlled laboratory conditions.

Firstly, the experimental results indicated that collision risk increased with rising levels of fatigue. In the conventional mechanical operation scenario, no collision risk was observed under low fatigue, low risk was associated with moderate fatigue, and high fatigue corresponded to high risk. In the HMI operation scenario, high fatigue corresponded to low risk, whereas low- and moderate-fatigue conditions did not trigger collision risk. Across both scenarios, increasing physical fatigue was generally accompanied by higher collision risk indices. Specifically, in the conventional mechanical operation scenario, elevated fatigue levels led to reduced safety distances and increased alarm frequency; similar trends were observed in the HMI scenario, although the degree of risk exposure differed between operation modes. These findings suggest that physical fatigue may adversely affect spatial control ability and risk exposure during tower crane operations under the experimental conditions.

Secondly, comparative analysis of the two operation scenarios under laboratory-simulated conditions revealed that, at moderate and high fatigue levels, the HMI scenario maintained relatively greater safety distances and lower alarm frequencies. Risk stratification results showed that, under moderate fatigue, the conventional mechanical operation scenario exhibited low risk, while no risk was triggered in the HMI scenario; under high fatigue, conventional operation corresponded to a high-risk range, whereas HMI operation remained within a lower-risk range. These results indicate that, in a controlled experimental environment, HMI may provide a moderating effect on risk exposure under fatigued conditions. However, it is important to emphasize that these findings reflect differences in operator performance under laboratory simulation and cannot be directly equated with a reduction in actual construction site accidents; further validation in real-world settings is required.

Overall, this study observed a trend linking physical fatigue to collision risk indices in a controlled experimental environment and preliminarily revealed the potential moderating role of HMI under fatigue. Given that the experimental platform simplifies the complexity of real construction environments, these conclusions should be regarded as exploratory findings rather than direct inferences to real-world operations.

This study also has several limitations that warrant further investigation. First, it primarily focused on the operational-level mechanisms of risk formation, analyzing the effects of physical fatigue and HMI on collision risk, without incorporating complex factors present on construction sites, such as organizational pressure, team coordination, management systems, or multitasking. Therefore, the results should be interpreted as trend observations under controlled experimental conditions rather than quantitative predictions of real-world risk levels. Second, although the experimental scenarios were designed to approximate the task workflow and risk structure of actual tower crane operations, the laboratory environment cannot fully replicate the spatial complexity and dynamic disturbances of real construction sites. Consequently, these conclusions should be cautiously evaluated before engineering application. Third, this study focused on the impact of physical fatigue on operational safety and did not integrate cognitive factors such as mental fatigue, sustained attention decline, or emotional stress. In real construction environments, multidimensional fatigue factors may interact, and their combined influence requires further investigation. Finally, the sample size was relatively limited; although a repeated-measures design and linear mixed-effects models were employed to control for individual differences, further stratified analysis based on operator experience or habitual practices was not conducted. Future research should consider on-site or semi-physical experiments in actual construction settings to enhance the external validity of the findings. Moreover, integrating multimodal physiological and behavioral data could enable more refined fatigue modeling and support the development of adaptive HMI strategies based on real-time fatigue recognition, providing a more reliable empirical foundation for the optimization of intelligent tower crane systems.